Re: [RFC/PATCHv2 v2 2/4] dma-mapping: Add dma_remap() APIs

[Laura Abbott <labbott@redhat.com>]

On 04/27/2016 08:25 AM, Catalin Marinas wrote:
> On Fri, Apr 22, 2016 at 05:35:16PM -0700, Stephen Boyd wrote:
>> Quoting Catalin Marinas (2016-04-21 03:35:12)
>>> On Tue, Apr 19, 2016 at 06:04:27PM -0700, Stephen Boyd wrote:
>>>> From: Laura Abbott <lauraa@codeaurora.org>
>>>>
>>>> Some systems are memory constrained but they need to load very
>>>> large firmwares. The firmware subsystem allows drivers to request
>>>> this firmware be loaded from the filesystem, but this requires
>>>> that the entire firmware be loaded into kernel memory first
>>>> before it's provided to the driver. This can lead to a situation
>>>> where we map the firmware twice, once to load the firmware into
>>>> kernel memory and once to copy the firmware into the final
>>>> resting place.
>>>>
>>>> This design creates needless memory pressure and delays loading
>>>> because we have to copy from kernel memory to somewhere else.
>>>> Let's add a couple DMA APIs that allow us to map DMA buffers into
>>>> the CPU's address space in arbitrary sizes. With this API, we can
>>>> allocate a DMA buffer with DMA_ATTR_NO_KERNEL_MAPPING and move a
>>>> small mapping window across our large DMA buffer to load the
>>>> firmware directly into buffer.
>>>
>>> The first two patches in this series don't make sense to me. I don't
>>> understand what the memory pressure is: physical or virtual? Because
>>> they don't seem to address the former (the DMA buffer is allocated in
>>> full) while the latter doesn't need any addressing at all on arm64, we
>>> have plenty of VA space.
>>>
>>> Why do you even need the coherent DMA API? Can you use the streaming API
>>> (map_sg etc.) with a separately allocated buffer?
>>
>> Hmm I guess I need to add in the patches that show how this is used on
>> top of "no-map" DT reserved memory regions. There are some more patches
>> that allow us to assigned reserved memory regions with the "no-map"
>> attribute to devices and then allocate from those regions using the
>> coherent DMA APIs. In the downstream kernel it's called a removed dma
>> pool[1].
>>
>> So the plan is to wire that all up so that the device can have a
>> reserved chunk of memory for the firmware that doesn't exist in the
>> kernel's linear memory mappings. Once we have allocated the region, we
>> can map it into the kernel's view of memory for a short time so that we
>> can load the firmware into it (dma_remap part). Once that's over, we
>> want to destroy the mapping so that we 1) don't use any of the kernel's
>> virtual memory space (dma_unremap part) to back the buffer and 2) so
>> that the secure world can protect the memory from the non-secure world.
>
> Does the firmware already know about such memory? If yes, I presume the
> kernel would have to be told about it and won't try to map it in the
> linear mapping.
>
> At this point, wouldn't a combination of:
>
> 	dma_declare_coherent_memory()
> 	dma_alloc_from_coherent()
> 	dma_release_from_coherent()
> 	dma_release_declared_memory()
>
> work? The removed_alloc() implementation in the link you posted doesn't
> seem far from dma_alloc_from_coherent(). The releasing of the declared
> memory above would unmap the memory, so there are no VA mappings left.
>

The removed alloc was specifically written as a fork of the coherent
pool. This was a choice for ease of out of tree maintenance. The better
choice here would be to fold those features back into dma-coherent.c
if needed.

Thanks,
Laura